Finite vector spaces and certain lattices

Finite vector spaces and certain lattices Thomas W. Cusick 106 Diefendorf Hall, Department of Mathematics, State University of New York at Buffalo, Buffalo, NY 14214-3093 E-mail: cusick@acsu.buffalo.edu Submitted: January 6, 1998; Accepted: March 18, 1998 Abstract The Galois number G n (q) is defined to be the number of subspaces of the n-dimensional vector space over the finite field GF (q). When q is prime, we prove that G n (q) is equal to the number L n (q) ofn-dimensional mod q lattices, which are defined to be lattices (that is, discrete additive subgroups of n-space) contained in the integer lattice Z n and having the property that given any point P in the lattice, all points of Z n which are congruent to P mod q are also in the lattice. For each n, we prove that L n (q) is a multiplicative function of q. Keywords: Multiplicative function; Lattice; Galois numbers; Vector space; Identities 1991 Mathematical Reviews subject numbers: Primary 05A15 05A19 11A25 11H06 Secondary 05A30 94A60 11T99

the electronic journal of combinatorics 5 (1998), #R17 2 1 Introduction The well known Gaussian coefficient (or q-binomial coefficient) ( ) n r q = (qn 1)(qn 1 1) (qn r+1 1) (q r 1)(q r 1 1) (q 1) is equal to the number of r-dimensional vector subspaces of the n-dimensional vector space V n (q) over the finite field GF (q). We let G n = G n (q) denote the total number of vector subspaces of V n (q). The numbers G n were named the Galois numbers by Goldman and Rota [4, p. 77]. Goldman and Rota [4] proved the recursion formula G n+1 =2G n +(q n 1)G n 1 (1) for the Galois numbers. Nijenhuis, Solow and Wilf [4] gave a different proof of (1) by using the observation that the r-dimensional vector subspaces of V n (q) are in one-to-one correspondence with the n by n matrices over GF (q) which have rank r and are in reduced row echelon form (rref). Recall that such a matrix is in rref if its last n r rows are all zeros; in each of the first r rows the first nonzero entry is a 1; the index of the i-th column (called a pivotal column) in which one of these r 1 s occurs strictly increases as i increases; and each of these r pivotal columns has only a single nonzero entry. We let E(r, n, q) denote the number of n by n matrices with rank r over the field GF (q) which are in rref. Then it was proved in [4] that G n (q) = n E(r, n, q). (2) r=0 The correspondence mentioned above gives E(r, n, q) = ( ) n r q. (3) For example, E(r, 4, 2) for r =0,1,2,3,4is1,15, 35, 15 and 1, respectively. We shall need the concept of an n-dimensional mod q lattice, which is defined to be an n-dimensional lattice contained in the integer lattice Z n and having the special property that given any point P in the lattice, all points of Z n which are congruent to P mod q are also in the lattice. Later in this paper we shall show how the mod q lattices are connected to the Galois numbers G n (q). It also turns out that the mod q lattices have an important application in cryptography, which we discuss elsewhere [2]. The set of mod q lattices contains various special subsets which can be used in the design of a novel kind of public-key cryptosystem. This idea originated with Ajtai [1].

the electronic journal of combinatorics 5 (1998), #R17 3 2 The multiplicative property We let L m (q) denote the number of m-dimensional mod q lattices. Our first goal is to prove that L m (q) is a multiplicative function, that is, for any positive integers r and s with gcd(r, s) =1wehaveL m (rs) =L m (r)l m (s). Theorem 1. The function L m (q) is multiplicative for each m =2,3,... Proof. Clearly, every m-dimensional mod q lattice is the solution space of some system Ax 0modq, (4) where A is an m by m matrix over the integers mod q. Conversely, the solution space of any system (4) is a mod q lattice. (Note that if e 1, e 2,...,e m is the standard basis for R m, then the m linearly independent vectors qe i (1 i m) are always solutions of (4), so the solution space is always a lattice of dimension m.) If gcd(r, s) = 1, there is a bijection between the set of m-dimensional mod rs lattices and the set of pairs of m-dimensional lattices made up of one mod r lattice and one mod s lattice. The bijection is defined as follows: Given a mod rs lattice which is the solution space of Ax 0modrs, we associate with it the pair of lattices which are solution spaces of where the matrices B and C are defined by Bx 0modrand Cx 0mods, (5) A B mod r and A C mod s; (6) and conversely, given (5) we define a matrix A by (6). To prove that this is a bijection, we must first show that different lattice pairs give different mod rs lattices. Given relatively prime integers r and s, by the definition of L m (q) we can choose two sets of matrices {B i :1 i L m (r)},whereb i is defined over the integers mod r, and{c i :1 i L m (s)},wherec i is defined over the integers mod s, such that every m-dimensional mod r lattice is the solution space of exactly one of the systems B i x 0modr, 1 i L m (r), and every m-dimensional mod s lattice is the solution space of exactly one of the systems C j x 0mods, 1 j L m (s). Since gcd(r, s) = 1, the theory of linear congruences in one variable shows that each pair of simultaneous congruences A B i mod r, A C j mod s, 1 i L m (r), 1 j L m (s) (7) defines a unique m by m matrix A = A ij, say, over the integers mod rs, andthese matrices are all different since the pairs B i,c j are. We shall show that the solution spaces (which are the mod rs lattices) of the systems are all distinct. A ij x 0modrs, 1 i L m (r), 1 j L m (s)

the electronic journal of combinatorics 5 (1998), #R17 4 Let A IJ and A KL be any two different matrices chosen from the A ij s. Then by (7), and {x mod r : A IJ x 0modrs} = {x : B I x 0modr} {x mod s : A IJ x 0modrs} = {x : C J x 0mods}; similar equations hold for A KL. Since the pairs B I,C J and B K,C L are different, we have either or {x : B I x 0modr} {x:b K x 0modr} {x : C J x 0mods} {x:c L x 0mods}, so the solution spaces for A IJ and A KL are different. Finally we must show that different mod rs lattices give different lattice pairs. This is clear since each congruence Ax 0modrs gives a unique pair of congruences (5), where the matrices B and C are defined by (6). 3 Counting mod q lattices Our first goal is to prove explicit formulas for the number of m-dimensional mod q lattices, which we denote by L m (q), when m is small. Theorem 2. The numbers L 2 (q) and L 3 (q) are given by L 2 (q) = ) gcd (k 1, qk2 k 1 q k 2 q (8) and L 3 (q) = k 1 q ) ) ) gcd (k 1, qk3 gcd (k 2, qk3 gcd (k 1, qk2. (9) k 2 q k 3 q We shall prove formula (8) first. We fix an x 1,x 2 Cartesian coordinate system in R 2. Given any 2-dimensional mod q lattice Λ, we have a basis-free representation for it as follows: The x 1 axis contains infinitely many points of Λ, with a density 1/k 1,wherek 1 is a positive integer which divides q. Every line x 2 = c either contains no points of Λ or contains a shifted copy of the set of lattice points on x 2 =0. If x 2 =k 2 is the line x 2 = c>0 which is closest to the x 1 axis and has points of Λ, then k 2 is a divisor of q. A line x 2 = c contains points of Λ if and only if has the form x 2 = tk 2 for some integer t. We say that Λ has jump k 2 (in the x 2 direction). If we

the electronic journal of combinatorics 5 (1998), #R17 5 let C 2 (Λ) denote the 2-dimensional volume of a fundamental cell of Λ, then we have C 2 (Λ) = k 1 k 2. To count the 2-dimensional mod q lattices which have given values of k 1 and k 2, it suffices to count the number of distinct 1-dimensional sublattices on x 2 = k 2 which give a mod q lattice. We define the shift s, wheresis an integer such that 0 s<k 1, to be the amount by which the 1-dimensional sublattice on x 2 = k is shifted with respect to the 1-dimensional sublattice on x 2 = 0. In order to give a mod q lattice, the shift s must give a 1-dimensional sublattice on x 2 = q which is an unshifted copy of the same sublattice on x 2 = 0. The sublattice on x 2 = q is shifted from the one on x 2 =0byqs/k 2,sotheshiftsgives a mod q lattice if and only if k 1 divides qs/k 2. (10) Clearly (10) holds for given k 1 and k 2 if and only if k 1 k 2 / gcd(k 1 k 2,q)=D,say, divides s. Thus there are k 1 /D =gcd(k 1,q/k 2 ) allowable values of s in the range 0 s<k 1. This proves (8). Now we prove formula (9). Each 3-dimensional mod q lattice Λ is made up of a 2-dimensional mod q sublattice in the x 1,x 2 plane, which we denote by P 0,and shifted copies of this sublattice in each of various planes P i (i nonzero integer) which are equally spaced parallel to P 0. As before, we let 1/k 1 denote the density of the points of Λ on the x 1 axis and we let k 2 denote the jump in the x 2 direction for the sublattice in P 0 (and so for Λ). The plane P 1 nearest to P 0 is at a distance k 3, where k 3 is a divisor of q. We say that Λ has jump k 3 in the x 3 direction. If we let C 3 (Λ) denote the 3-dimensional volume of a fundamental cell of Λ, then we have C 3 (Λ) = k 1 k 2 k 3. To count the 3-dimensional mod q lattices with given k 1,k 2 and k 3, for each 2- dimensional mod q sublattice on P 0 we count the number of distinct 2-dimensional sublattices in x 3 = k 3 (i.e., the plane P 1 )whichgiveamodqlattice. We let s denote the shift for the 1-dimensional sublattices in P 0, as before, and we define the (vector) shift s =(s 1,s 2 ), where 0 s i <k i (i=1,2), to be the amount by which 0 in P 0 is moved when we go to the sublattice in P 1. The shift s gives a mod q lattice if and only if k 1 divides qs 1 /k 3 and k 2 divides qs 2 /k 3, (11) that is, if and only if the orthogonal projection of (q/k 3 )(s 1,s 2,k 3 ) into the plane P 0 is a lattice point. Now (11) holds for given k 1,k 2 and k 3 if and only if k i k 3 / gcd(k i k 3,q)= D i, say, divides s i (i =1,2). Thus there are k i /D i =gcd(k i,q/k 3 ) allowable values of s i in the range 0 s i <k i.thisproves(9). It is possible to extend the formula in Theorem 2 to the case of general m, but complicated m-fold sums are involved. Since we do not need this result, we do not give it here. A multiplicative function is completely determined by its values at prime powers, so it is of interest to examine L m (p a ) for prime p. Direct calculation using (8) gives L 2 (p a )= a (1 + 2i)p a i = (p +1)pa+1 (2a +3)p+2a+1. (p 1) 2 i=0

the electronic journal of combinatorics 5 (1998), #R17 6 Computer calculations using (9) give Table 1, which shows the expansion of L 3 (p a ) in powers of p for small a. There does not seem to be any nice explicit formula for L 3 (p a ), though various properties of the coefficients in the table can be deduced. Table 2 gives some values for L 2 (q) andl 3 (q). a, j 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 4 2 2 2 7 6 6 5 3 3 10 10 12 10 10 8 4 4 13 14 18 17 18 14 15 11 5 5 16 18 24 24 28 22 24 20 20 14 6 6 19 22 30 31 38 32 35 30 30 27 25 17 7 7 22 26 36 38 48 42 48 42 42 38 38 34 30 20 8 Table 1: Coefficients of p j in the expansion of L 3 (p a ),a 7. 2 3 4 5 7 8 9 11 13 16 17 19 23 L 2 (q) 5 6 15 8 10 37 23 14 16 83 20 22 26 L 3 (q) 16 28 131 64 116 830 457 268 368 4633 616 1016 1108 Table 2: Values of L 2 (q) andl 3 (q) for small prime powers q. 4 The connection with Galois numbers Because of (2), our next theorem shows that L m (q) =G m (q) whenever q is a prime. Theorem 3. For any prime q, we have L m (q) = m E(r, m, q). r=0 Proof. We have already seen that every m-dimensional mod q lattice is the solution space of some system (4), where A is an m by m matrix over the integers mod q. Conversely, the solution space of any system (4) is an m-dimensional mod q lattice. Since q is prime, the mod q lattices are thus in one-to-one correspondence with the m by m reduced row echelon forms of matrices over GF (q) and we have the desired equation. Because of (3), it is easy to compute E(r, m, q) for given values of r, m, q. If q is not prime, the first two sentences in the proof of Theorem 3 are still true, so the one-to-one correspondence between the mod q lattices and solution spaces of systems (4) is still valid. What is lost is the link with matrices over a field which

the electronic journal of combinatorics 5 (1998), #R17 7 are in reduced row echelon form (rref). Thus this paper shows that there are two different natural extensions of the Galois numbers G n (q), q prime. One extension leads to the Galois numbers G n (q) for arbitrary positive integers q, as given in [4]. In that paper a formal definition of a rref matrix over a set of q symbols is given and finite fields play no role. For each n, thenumbersg n (q) are fixed polynomials in q, and the recursion (1) holds as a polynomial identity. The other extension leads to the multiplicative functions L n (q) in this paper. If q is not prime, then L n (q) isnot a polynomial in q and the analog of (1) does not hold. References [1] Miklos Ajtai, Generating hard instances of lattice problems, in: Proc. 28th ACM Symposium on the Theory of Computing, 1996, pp. 99-108. [2] Thomas W. Cusick, The Ajtai random class of lattices, to appear. [3] Jay Goldman and Gian-Carlo Rota, The number of subspaces of a vector space, in: Recent Progress in Combinatorics, ed. W. T. Tutte (Academic Press, 1969), pp. 75-83. [4] Albert Nijenhuis, Anita E. Solow and Herbert S. Wilf, Bijective methods in the theory of finite vector spaces, J. Combin. Theory (A) 37 (1984), 80-84.